IndicTrans2 indic→en (distilled 200M) — int8 ONNX (for ChatPanel)

An int8 ONNX export of ai4bharat/indictrans2-indic-en-dist-200M by AI4Bharat (MIT licence, LICENSE), for ChatPanel's local gateway to translate meeting lines in Indian languages into English on the person's own computer with onnxruntime and no Python.

File What
encoder.onnx + .data ids (1, S) → states (1, S, 512)
cross_kv.onnx + .data cross-attention keys/values for the 18 decoder layers
decoder.onnx + .data one cached decoder step → logits
src.json, tgt.json source BPE pieces + scores and vocabulary; target vocabulary

The text preparation of AI4Bharat's IndicTransToolkit (normalisation, script unification to Devanagari, language tags, SentencePiece) is ported to JavaScript in ChatPanel's gateway. Checked against the original on FLEURS dev: preprocessing 49/50 identical (the one difference is the toolkit emitting a literal ़), pieces and ids 50/50, greedy translations 11/12 identical, ~95 ms per sentence on four CPU threads; chrF against FLEURS English, Telugu 68.9 and Hindi 65.1 (opus-mt-mul-en: 51.5 and 43.6).

Please cite IndicTrans2: Gala et al., IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages, TMLR 2023.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ChatPanel/indictrans2-indic-en-int8

Quantized
(7)
this model